We process and visualize Swedish parliamentary data using methods from statistics and machine learning, which allows us to obtain insight into the political processes behind the data. We produce plots that let us infer the relative stance of political parties and their members on different topics. In addition, we can infer the degree of homogeneity of individual votes within different parties, as well as the degree of multi-dimensionality of Swedish politics. With a changing political landscape and new political movements gaining traction across the Nordic countries, Europe and the world, it is especially important for political scientists and citizens to be able to assess the political stance of political actors with respect to specific topics. Using methods from statistics and machine learning on Swedish parliamentary data, we obtain plots that allow us to visualize the relative stance of political parties and members of the Swedish Parliament1 (MSPs) on different topics, the degree of homogeneity of individual votes within different parties, as well as the degree of multi-dimensionality of Swedish politics. Although the Swedish political landscape is an especially interesting case due to its relatively high number of parties, the methods can be generalized to any open parliamentary data set that includes voting records. Section 2 explains how the data was acquired and pre-processed. Section 3 presents the results of our analysis. Section 4 presents lines of work through which we are currently working or plan to work.
We use parliamentary data publicly available at the website of the Parliament of Sweden2 covering the period from 2014-11-19 to 2018-06-20.
1 In order to avoid confusion of names, in this paper, we use the acronym MSP (for “member of the Swedish Parliament”), rather than the expected and more common MP (for “member of parliament”), since the latter acronym is long established in Swedish politics as the conventional abbreviated reference to Miljöpartiet de Gröna, the Swedish Green party. 2 http://data.riksdagen.se
The data is originally obtained from files in comma-separated values (CSV) format. Each file contains data for one parliamentary working year, organized in lines of text, each containing a fixed number of text fields separated by commas. Each of the fields contains a specific value for a variable: identifier of a voting session, identifier of the document about which the voting session is about, name of an MSP, value of the vote of the MSP in that voting session, date of the voting session, etc. Some additional information is included about MSPs, such as which party they are member of, their birth date, gender, and constituency. In our current experiments we only use the party affiliation. Each document identifier contains the identifier of one of the 15 existing parliamentary committees,3 which define topics such as taxation, finance, health and welfare, etc. We extract this, which allows us to connect each voting session to one of these topics.
We carry out some preprocessing to normalize the data. The names of one party changes in this period and we use the latter name only (Folkpartiet ! Liberalerna). The names of the MSPs are stored in the CSV files in different formats, sometimes given-name first and in other cases family-name first. We attempt to unify this by converting the first format into the second one. If an MSP leaves a party, we use their last affiliation. The Committee on Defense and the Committee on Foreign Affairs are joined into one because some voting sessions use a joint identifier. There is one voting session using a joint identifier for the Committee on Cultural Affairs and the Committee on Civil Affairs; since there seems to be little topical overlap between these two, we chose to omit that voting session rather than conflating all the other voting sessions of the two committees. The value of the vote of an MSP in a voting session can be ‘Yes’, ‘No’, ‘Abstention’, ‘Absence’ and ‘Nonexistence’. ‘Nonexistence’ reflects reflects the situation where the individual in question was not an MSP at the time of the voting, whereas ‘Absence’ reflects that the individual was an MSP but was not present during the voting session. ‘Nonexistence’ is the only value not explicitly included as a record in the CSV files, so it is inferred for each voting session as the voting value for the MSPs for which there is an explicit voting value in other voting sessions in the time period under analysis, but not in the voting session at hand. In the remainder of the document, we will refer to ‘Yes’, ‘No’, ‘Abstention’ and ‘Absence’ as ‘existing votes’. We encode ‘Yes’ as a numerical value +1, ‘No’ as 1 and both ‘Abstention’, and ‘Absence’ as a 0. In some experiments where a numerical value is required for every MSP and every voting session, we encode ‘Nonexistence’ as 0 as well. However, because ‘Nonexistence’ does not reflect an actual choice of the MSP, assuming its equivalence to abstention and absence introduces a kind of noise – had that individual been an MSP at that time, the vote might as well have been ‘Yes’ or ‘No’. For this reason, in order to prevent this noise from distorting our conclusions, experiments that reflect information about MSPs will differentiate between ‘Continuous MSPs’, defined as those MSPs for which there is an existing vote for each voting session in the period analyzed (2010–2018); and ‘Discontinuous MSPs’, defined 3 These are explained at
The-15-parliamentary-committees/ https://www.riksdagen.se/en/Committees/ 378 as those MSPs for which there is at least one voting session for which there is no existing vote in the period analyzed.
This results in information about 427 MSPs, 267 of which are continuous. In the 2010– 2018 period there are 2,583 voting sessions, each of which contains exactly 349 valid votes,4 totaling 901,467 valid votes in the whole period. Each voting session is associated to one among 14 committees/topics.
Table 1 shows the parties in the period of time analyzed, together with the number of distinct MSPs throughout that period. The abbreviations and colors in the table will be used in the following sections. We obtain three different types of views on the parliamentary data. Each view is based on different mathematical and computational methods and aims at giving insight into different political aspects reflected in the parliamentary data.
– We build similarity charts that show how close parties are to other parties, and MSPs
to other MSPs, in terms of their voting record. This is done considering all votes as
a whole, as well as individually for votes associated to specific topics defined by the
parliamentary committees. We apply agglomerative hierarchical clustering
Figure 1 displays charts showing degrees of similarity (obtained through cosine distance)
between the votes of each party (which are obtained averaging the votes of the MSPs on each
party). Each chart is obtained using the existing votes associated to each committee, and one
extra chart is obtained using all votes. Using single-linkage hierarchical clustering
Similarly, Figure 2 contains a chart showing degrees of similarity (obtained through cosine distance) between the votes of each continuous MSP. Discontinuous MSPs are omitted since any attempt at guessing a non-existent vote for a voting session would introduce a bias. Voting sessions from all committees are considered. MSPs are again ordered using singlelinkage hierarchical clustering, which causes those with similar voting records to be close in the linear order, and the block structure to become clearly visible.
In Figures 1 and 2, the three most visible blocks correspond to the well-known formal blocks in Swedish politics: the Alliance (M+L+C+KD), the Red-Greens (S+MP+V) and the Swedish Democrats (SD and independents, all of which belonged previously to SD). In addition to this, one can draw additional conclusions. Within the Red-Greens, V stands apart from S and MP, which are very close to each other on average. In Figure 2, the clustering algorithm is not even able to group members of MP in a single cluster separate from S, which means that the differences among the votes of members of MP are comparable to those between them and the members of S. The division by topics in Figure 1 shows additional insights: for instance, on constitutional affairs, the Alliance and the Red-Greens form a single block opposed to SD. The difference between V and S+MP is lower in some committees such as Finance and Taxation, which should correspond to a political agreement on such issues. The gap between SD and S+MP is comparatively lower in some committees such as Civil Affairs Fig. 1: Charts showing similarity between the votes of each party (which are obtained averaging the votes of the MSPs of each party). Each chart is obtained using the votes associated to each committee, and one extra chart is obtained using all votes. Higher levels of similarity are shown in darker shades. The diagonals are black because they stand for the perfect similarity of a party with itself. Fig. 2: Similarity (obtained through cosine distance) between the votes of each continuous MSP. All votes are considered. Higher levels of similarity are shown in darker shades. The diagonal is black because it stands for the perfect similarity of an MSP with itself. The side colored bars represent the party of each MSP. The individual names of the 160 continuous MSPs are omitted due to space constraints. The legend for colors can be found in Table 1 and Figure 4. 382 and Industry and Trade. On ‘Finance’, M stands farther from the Red-Green block as compared to the rest of the Alliance, which produces a strong sub-block. On ‘Social Insurance’, L stands closer to the Red-Green block as compared to the rest of the Alliance. On ‘Environment and Agriculture’, L stands closer to the Red-Green block as compared to the rest of the Alliance. MSPs classified as independent are very close to SD on all matters, which is not surprising since all of them happen to be former members of SD (and their votes during their period in SD are counted as well even if they are labeled as independents). 3.2
When dealing with high-dimensional data, one cannot visualize more than two or at most
three dimensions in a single plot (using X,Y, and Z axes). The voting data is high-dimensional
because each MSP is a point with several thousands of dimensions (one for each voting
session). Principal Component Analysis
Figure 4 shows MSPs plotted in two-dimensional subspaces defined by pairs of different
principal components. In each plot, the horizontal and vertical axes represent each two
principal components. The clustering of MSPs according to party is clear. It also shows which
MSPs have voting records closer to those from other parties. Additional components may
reflect different groupings, but the higher the component index, the smaller are the distances
reflected in their axes. The plot is akin to a multi-dimensional political compass
Figure 5 shows how much spread (variance) is contained in each of the principal components. It shows that the voting data cannot be linearly squeezed into a straight line (like the traditional left-right axis), although a 5-dimensional space constitutes a reasonable summary of the data. 4
We are currently working on extending the work presented in this paper in the following ways.
– Currently, the topic of each vote is determined by the committee assigned to the vote.
However, some committees conflate topics that might be relevant to analyze
independently (for instance, the topics of “citizenship and migration” and “national pensions”
are arguably distinct but both handled by the Committee on Social Insurance), while,
hypothetically, the same topic may be addressed with different perspectives across several
committees. Each voting session has an associated document from which topics could
be induced and we are comparing different language technology methods in order to
automatically assign topics to voting sessions.
– Currently, the axes of the political compass in Figure 4 are unlabeled and, therefore, lack
a clear interpretation. We plan to apply language technology techniques on the
parliamentary documents associated to the voting sessions in order to identify the topics associated
to each axis. We are also testing other matrix factorization methods that may be more
appropriate for the task, like Independent Component Analysis
This work has been supported by a framework grant (Towards a knowledge-based culturomics;5 contract 2012-5738) as well as funding to Swedish CLARIN (Swe-Clarin;6 contract 2013-2003), both awarded by the Swedish Research Council, and by infrastructure funding granted to Språkbanken by the University of Gothenburg.
5 https://spraakbanken.gu.se/eng/culturomics 6 https://sweclarin.se/eng